[Ansor][AutoTVM v2.0] Phase 1: feature extraction for cost models (ap…

…ache#6190)

* [AutoScheduler] add feature extraction

* fix lint

* fix gpu test

* address comments

* improve flop estimation

* rebase

* refactor with group

* fix

* Apply suggestions from code review

include/tvm/auto_scheduler/feature.h

@@ -0,0 +1,122 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied.  See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ * \file auto_scheduler/feature.h
+ * \brief Feature extraction for the cost model.
+ * We extract one feature vector per BufferStoreNode statement in a TIR Stmt,
+ * so we call this feature as "Per Store" feature.
+ * The cost model also does prediction for each BufferStoreNode statement and aggregates
+ * the predictions as the whole score for a TVM IR (Stmt).
+ *
+ * The feature specification is defined by `src/auto_scheduler/feature.cc:: FeatureSet`
+ */
+
+#ifndef TVM_AUTO_SCHEDULER_FEATURE_H_
+#define TVM_AUTO_SCHEDULER_FEATURE_H_
+
+#include <tvm/auto_scheduler/compute_dag.h>
+#include <tvm/auto_scheduler/measure.h>
+
+#include <string>
+#include <vector>
+
+namespace tvm {
+namespace auto_scheduler {
+
+/*!
+ * \brief Get per-store feature from a TIR Stmt
+ * \param stmt The input lowered TIR statement
+ * \param cache_line_size The size of cache line in bytes
+ * \param max_n_bufs The maximum number of extracted buffers for one statement
+ * \param ret The returned feature vector
+ */
+void GetPerStoreFeature(const Stmt& stmt, int cache_line_size, int max_n_bufs,
+                        std::vector<float>* ret);
+
+/*
+ * \brief Get the names of elements in the feature vector. Use this for debug and inspection.
+ * \param max_n_bufs The maximum number of extracted buffers for one statement
+ * \param ret The returned names.
+ */
+void GetPerStoreFeatureName(int max_n_bufs, std::vector<std::string>* ret);
+
+/*!
+ * \brief Get per-store feature from states of the same task
+ * \param states The input states
+ * \param task The same search task for all states
+ * \param skip_first_n_feature_extraction Skip feature extraction for the first n states
+ * \param max_n_bufs The maximum number of extracted buffers for one statement
+ * \param features The returned feature vector. The innermost vector contains the
+ * feature vectors for all BufferStoreNode statements
+ */
+void GetPerStoreFeaturesFromStates(const Array<State>& states, const SearchTask& task,
+                                   int skip_first_n_feature_extraction, int max_n_bufs,
+                                   std::vector<std::vector<float> >* features);
+
+/*!
+ * \brief Get per-store feature from states of different tasks
+ * \param states The input states
+ * \param tasks The search tasks corresponding to the input states
+ * \param skip_first_n_feature_extraction Skip feature extraction for the first n states
+ * \param max_n_bufs The maximum number of extracted buffers for one statement
+ * \param features The returned feature vector. The innermost vector contains the
+ * feature vectors for all BufferStoreNode statements
+ */
+void GetPerStoreFeaturesFromStates(const Array<State>& states, const std::vector<SearchTask>& tasks,
+                                   int skip_first_n_feature_extraction, int max_n_bufs,
+                                   std::vector<std::vector<float> >* features);
+
+/*!
+ * \brief Get per-store features from a log file
+ * \param filename The name of log file
+ * \param max_lines Only read the first n lines of the file
+ * \param max_n_bufs The maximum number of extracted buffers for one statement
+ * \param features The returned feature vector. The innermost vector contains the
+ * feature vectors for all BufferStoreNode statements
+ * \param normalized_throughputs The normalized throughputs for all states
+ * \param task_ids The task ids for all states
+ */
+void GetPerStoreFeaturesFromFile(const std::string& filename, int max_lines, int max_n_bufs,
+                                 std::vector<std::vector<float> >* features,
+                                 std::vector<float>* normalized_throughputs,
+                                 std::vector<int>* task_ids);
+
+/*!
+ * \brief Get per-store features from measurement input/result pairs
+ * \param inputs The meaurement inputs
+ * \param results The measurement results
+ * \param skip_first_n_feature_extraction Skip feature extraction for the first n meaurement pairs
+ * \param max_n_bufs The maximum number of extracted buffers for one statement
+ * \param features The returned feature vector. The innermost vector contains the
+ * feature vectors for all BufferStoreNode statements
+ * \param normalized_throughputs The normalized throughputs for all states
+ * \param task_ids The task ids for all states
+ */
+void GetPerStoreFeaturesFromMeasurePairs(const Array<MeasureInput>& inputs,
+                                         const Array<MeasureResult>& results,
+                                         int skip_first_n_feature_extraction, int max_n_bufs,
+                                         std::vector<std::vector<float> >* features,
+                                         std::vector<float>* normalized_throughputs,
+                                         std::vector<int>* task_ids);
+
+}  // namespace auto_scheduler
+}  // namespace tvm
+
+#endif  // TVM_AUTO_SCHEDULER_FEATURE_H_

python/tvm/auto_scheduler/__init__.py

@@ -23,6 +23,7 @@
 from . import loop_state
 from . import utils
 from . import workload_registry
+from . import feature
 
 # Shortcut
 from .auto_schedule import SearchTask, TuningOptions, HardwareParams, \

python/tvm/auto_scheduler/feature.py

@@ -0,0 +1,242 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+
+""""
+Python API for Feature extraction. The extracted features vector are used by cost models.
+
+We extract one feature vector per BufferStoreNode statement in a TIR Stmt,
+so we call this feature as "Per Store" feature.
+The cost model also does prediction for each BufferStoreNode statement and aggregates
+the predicted score of each BufferStoreNode as the score of a TIR Stmt.
+
+The feature specification is defined by `src/auto_scheduler/feature.cc::FeatureSet`
+"""
+
+from typing import List, Tuple, Union, Optional
+import struct
+
+import numpy as np
+
+from .loop_state import State, StateObject
+from .measure import MeasureInput, MeasureResult
+from . import _ffi_api
+
+# The maximum number of extracted buffers for one statement
+DEFAULT_MAX_N_BUFS = 5
+
+# The length of the feature vector
+DEFAULT_FEATURE_VEC_LEN = 164
+
+# The size of int and float in bytes
+SIZE_OF_INT32 = 4
+SIZE_OF_FLOAT32 = 4
+
+def unpack_feature(byte_arr: bytearray) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+    """Unpack the flatten feature (in byte array format) from c++
+
+    Parameters
+    ----------
+    byte_arr: bytearray
+        The two-dimensional feature vector in serialized byte array format
+
+    Returns
+    -------
+    features: np.ndarray
+        Feature vectors
+    normalized_throughputs: np.ndarray
+        Normalized throughputs
+    task_ids: np.ndarray
+        Task ids
+    """
+
+    # The format for n records is:
+    # {
+    #   int n;
+    #   int[n+2] sizes
+
+    #   float[sizes[0]]    feature for record 1
+    #   float[sizes[1]]    feature for record 2
+    #   ...                feature for record i...
+    #   float[sizes[n-1]]  feature for record n
+
+    #   float[sizes[n]]    normalized throughput for n records
+    #   int[sizes[n+1]]    task id for n records
+    # }
+
+    vec_len = DEFAULT_FEATURE_VEC_LEN
+
+    # unpack sizes
+    offset = 0
+    n = struct.unpack_from("1i", byte_arr, offset=offset)[0]
+    offset += SIZE_OF_INT32
+
+    sizes = struct.unpack_from("%di" % (n+2), byte_arr, offset=offset)
+    offset += SIZE_OF_INT32 * (n+2)
+
+    # unpack features
+    features = []
+    for size in sizes[:-2]:
+        row = []
+
+        # Now, we need to unpack the feature for multiple statements.
+        # The format is:
+        # {
+        #     int n_stmts
+        #     float[n_stmt][vec_len] feature_vecs
+        # }
+        # where vec_len can be calculated by `(size - 1) / n_stmts`
+
+        if size == 0:
+            # failed during lowering
+            features.append(np.zeros((1, vec_len)))
+        else:
+            n_stmts = struct.unpack_from("f", byte_arr, offset=offset)
+            offset += SIZE_OF_FLOAT32
+
+            n_stmts = int(n_stmts[0] + 0.5)
+            tmp_vec_len = (size - 1) // n_stmts
+            assert tmp_vec_len == vec_len, "The lenght of feature vector is wrong. " \
+                                           "Expected %d but got %d." % (vec_len, tmp_vec_len)
+            assert tmp_vec_len * n_stmts == size - 1
+            for _ in range(n_stmts):
+                x = struct.unpack_from("%df" % vec_len, byte_arr, offset=offset)
+                offset += vec_len * SIZE_OF_FLOAT32
+                row.append(x)
+
+            features.append(np.array(row))
+
+    # unpack normalized_throughputs
+    m = sizes[-2]
+    normalized_throughputs = struct.unpack_from("%df" % m, byte_arr, offset=offset)
+    offset += m * SIZE_OF_INT32
+
+    # unpack task_ids
+    m = sizes[-1]
+    task_ids = struct.unpack_from("%di" % m, byte_arr, offset=offset)
+    offset += m * SIZE_OF_INT32
+
+    assert offset == len(byte_arr), "%d vs %d" % (offset, len(byte_arr))
+    return np.array(features, dtype=object), np.array(normalized_throughputs), np.array(task_ids)
+
+
+def get_per_store_features_from_file(filename: str,
+                                     max_lines: int,
+                                     max_n_bufs: Optional[int] = None) \
+        -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+    """Get per_store features from a log file
+
+    Parameters
+    ----------
+    filename: str
+        The input filename
+    max_lines: int
+        Only extract the first n lines of the file
+    max_n_bufs: Optional[int]
+        The maximum number of extracted buffers for one statement
+
+    Returns
+    -------
+    features: np.ndarray
+        Feature vectors
+    normalized_throughputs: np.ndarray
+        Normalized throughputs
+    task_ids: np.ndarray
+        Task ids
+    """
+    byte_arr = _ffi_api.GetPerStoreFeaturesFromFile(
+        filename, max_lines, max_n_bufs or DEFAULT_MAX_N_BUFS)
+    return unpack_feature(byte_arr)
+
+
+def get_per_store_features_from_measure_pairs(inputs: List[MeasureInput],
+                                              results: List[MeasureResult],
+                                              skip_first_n_feature_extraction: int = 0,
+                                              max_n_bufs: Optional[int] = None) \
+        -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+    """Get per_store features from measurement input/result pairs
+
+    Parameters
+    ----------
+    inputs: List[MeasureInput]
+        The measure inputs
+    results: List[MeasureResult]
+        The measure results
+    skip_first_n_feature_extraction: int
+        Skip feature extraction for the first n states
+    max_n_bufs: int
+        The maximum number of extracted buffers for one statement
+
+    Returns
+    -------
+    features: np.ndarray
+        Feature vectors
+    normalized_throughputs: np.ndarray
+        Normalized throughputs
+    task_ids: np.ndarray
+        Task ids
+    """
+    byte_arr = _ffi_api.GetPerStoreFeaturesFromMeasurePairs(
+        inputs, results, skip_first_n_feature_extraction, max_n_bufs or DEFAULT_MAX_N_BUFS)
+    return unpack_feature(byte_arr)
+
+
+def get_per_store_features_from_states(states: List[Union[State, StateObject]],
+                                       task: "SearchTask",
+                                       max_n_bufs: Optional[int] = None) -> List[np.ndarray]:
+    """Get per_store features from measurement input/result pairs
+
+    Parameters
+    ----------
+    states: List[Union[State, StateObject]]
+        The input states
+    task: SearchTask
+        The search task of the input states
+    max_n_bufs: Optional[int]
+        The maximum number of extracted buffers for one statement
+
+    Returns
+    -------
+    features: np.ndarray
+        Feature vectors
+    normalized_throughputs: np.ndarray
+        Normalized throughputs
+    task_ids: np.ndarray
+        Task ids
+    """
+    if isinstance(states[0], State):
+        state_objects = [s.state_object for s in states]
+    elif isinstance(states[0], StateObject):
+        state_objects = states
+    byte_arr = _ffi_api.GetPerStoreFeaturesFromStates(
+        state_objects, task, max_n_bufs or DEFAULT_MAX_N_BUFS)
+    return unpack_feature(byte_arr)[0]
+
+
+def get_per_store_feature_names(max_n_bufs: Optional[int] = None) -> List[str]:
+    """Get the name of every element in the feature vector. Use this for debug and inspection.
+
+    Parameters
+    ----------
+    max_n_bufs: int
+        The maximum number of extracted buffers for one statement
+
+    Returns
+    -------
+    names: List[str]
+        The names of elements in the flatten feature vector
+    """
+    return _ffi_api.GetPerStoreFeatureNames(max_n_bufs or DEFAULT_MAX_N_BUFS)